(navigation image)
Home American Libraries | Canadian Libraries | Universal Library | Community Texts | Project Gutenberg | Children's Library | Biodiversity Heritage Library | Additional Collections
Search: Advanced Search
Anonymous User (login or join us)
Upload
See other formats

Full text of "Speech commands in control systems"

NASA TECHNICAL TRANSLATION 



NASA TT F-11 ,252 



CM 
LTV 
CM 



< 
CO 

< 



SPEECH COMMANDS IN CONTROL SYSTEMS 
E. Kyunnap 



Translation of "Ustnyye Komandy v Sistemakh Upravleniya" 

Izvestiya Akademi i Nauk Estonskoy SSR, Seriya Fiziko- 

Matematicheskikh i Tekhnicheski kh Nauk, Vol. 15, No. 3, 

pp. 377-399, 1966 



/ 




(THRU) 



(CODE) 



(CATEGORY) 



07- 





NATIONAL AERONAUTICS AND SPACE ADMINISTRATION 
WASHINGTON, DC 20546 APRIL 1 968 



NASA TT F-l 1,252 



SPEECH COMMANDS IN CONTROL SYSTEMS 
E. Kyunnap 

ABSTRACT. A review of the literature dealing with auto- 
matic recognition of speech sounds. The problem of 
increasing channel carrying capacity is considered. A 
study is made of the mechanism of sound formation. The 
principles of operation of band-pass, formant, scanning, 
harmonic and correlation voice coders are outlined. A num- 
ber of signal devices for recognizing speech signals are 
described, and the use of universal computers as a means of 
studying and recognizing speech signals is discussed. 

1. Multiplexing of the Communications Channel 

1 
As is known, a speech signal consists of a sum of individual oscillations / 377 
of various frequencies and amplitudes. When it is expanded into a series, 
summation may be performed either with respect to elements equally spaced by 
frequency (Fourier series) or with respect to elements equally spaced by time 
(Kotel'nikov's theorem). In the first case, the speech signal is subjected to 
harmonic analysis and the amplitudes (and phases) of the harmonics are trans- 
mitted through the channel; at the receiving point, the speech signal is 
restored using these spectral coefficients, determined by the analyzer at the 
transmitting end of the communications channel. In the second case, pulses 
are transmitted through the communications channel at discrete time intervals; 
the amplitudes of the pulses are proportional to the instantaneous values of 
the function, read at intervals At. At the receiving end, these pulses pass 
through a filter, the output of which is constantly added. 

The information transmitted contains, in addition to the useful informa- 
tion, a certain quantity of noise. A criterion characterizing the signal level 
in comparison to the noise level is the value H = log p/p , where p and p are 

the mean powers of signal and noise respectively. The product of three 
quantities V = TFH is called the signal volume. A communications channel is 
also characterized by three quantities: T, , the time interval during which the 

channel is connected; F, , the band of frequencies transmitted through the 

channel; and H, , the power level of the apparatus making up the channel. The 

product V, = T, F,H, is called the channel capacity. The condition V, > V must 

be assured if the signal is to pass through the channel. 

1 Numbers in the margin indicate pagination in the foreign text. 



Multiplexing of a communications channel can be achieved by deforming one 
of the pairs of these quantities while leaving the third quantity unchanged. 
Deformation of the signal is performed by compressing it at the transmitting /378 
end of the channel and expanding it at the receiving end. Changes in H, T and 
F correspond to changes in amplification, delay of the signal using a delay 
line and frequency respectively. For example, if a speech signal is recorded 
on magnetic tape, transmitted at double speed, re-recorded at the receiving 
end and then played back at half speed, the volume of information transmitted 
remains unchanged, but transmission is performed twice as fast at twice the 
frequency. 

Companding of a speech signal, i.e. compression at the transmitting end 
and expansion at the receiving end of a communications channel, may be 
frequency, amplitude or time companding. One extreme form of amplitude 
companding of a speech signal is clipping. In this case, the amplitude of 
the speech signal is limited at two levels and only the points at which the 
function changes its sign are transmitted. 

As we know, the smaller the base of a number system, the greater the 
number of digits required to represent the same numbers. An optimal system 
would be a system with the base e. However, it is impossible in practice to 
produce such a system, and the base used must be either 2 or 3. The most 
widely used system is the binary system, although the trinary system would be 
more efficient, since 3 is closer than 2 to the value of e. Speech clipping 
(Figure 1) corresponds to a vinary number system of transmission. As 
I. Licklider has stated [111, 112], when a speech signal is limited to two 
levels, a sufficient quantity of information still remains in the signal to 
provide intelligibility at the required level. The technical conditions for 
impulse telephony call for 128 levels. Consequently, clipping a speech 
signal decreases its volume by a factor of 7. 

The intelligibility of speech increases if the speech signal is differ- 
entiated before clipping. With this technique, the frequency of the clipped 
signal is increased and the location of the extreme points of the original 
signal is transmitted. Partial companding is achieved by dividing the 
frequency of this signal at the transmitting end and correspondingly 
multiplying it at the receiving end [36, 156] . The spectrum is narrowed by 
only a factor of 6, and great distortion of the speech results; therefore, 
this companding is not particularly promising. However, reduction of the 
frequency range of speech by limiting the upper and lower ends of the spectrum 
has the opposite effect. The human voice occupies a frequency range from 
50-60 Hz to 15-20 KHz. If the only demand placed on a telephone conversation 
is intelligibility and recognition of the voice of the person speaking, the 
upper frequency limit can be reduced to 2.5-3 KHz. If intelligibility alone 
is required, the signal volume can be decreased still further. Under noise 
conditions, limiting the frequencies transmitted over a telephone channel to a 
maximum of 3500 Hz and a minimum of 300 Hz increases intelligibility of the 
speech [31, 130, 136]. 




Figure 1 . CI ipping of 
a Speeeh S ignal : 
a, Original speech 
signal ; b, CI ipped 
speech signal 



Time companding of speech eliminates 
certain time intervals, and the pauses which 
arise are filled with other transmitted 
material. At the receiving end, the pauses in 
the speech are filled in by the listener ment- 
ally [25, 36, 39, 73]. However, the low deg-ee 
of companding achieved (up to 2) and the 
reduction in intelligibility indicates that 
this method has no particular future prospects 
[103]. Better results for time multiplexing of 
communications channels are yielded by using 
the pauses between words and phrases in natu-al 
speech, as well as the pauses on one line while 
the conversation partner speaks on the other 
line. The switching of individual conversa- 
tions on one line into the pauses in other 
conversations can be performed by clipping the 
speech signal. With this system, four to six 
conversations can be transmitted through one 
communications channel. 



Parametric companding methods, although they allow considerably greater 
compression of a communications channel, disrupt the micro structure of the 
speech signal: only the parameters produced by a speech signal analyzer are 
transmitted through the channel, and at the receiving end these parameters are 
used to control a speech synthesizer. Thus, the parametric methods of 
companding involve automatic recognition and synthesis of speech signals. 



/379 



The devices used to transmit the speech signal by the parametric method 
have come to be called vocoders [59,60], In semi-vocoder devices, the para- 
metric method is used to transmit only the upper portion of the speech 
signal, and the lower portion is transmitted continually [69]. In addition to 
providing multiplexing of communications channels, vocoders can be used to 
provide secrecy for telephone conversations [2, 102, 157], 



Investigation of the Formation of Sound 



A number of works have been dedicated to the investigation of the 
functioning of the ear [14, 84, 92, 114] and the vocal cords [64, 89, 94, 
165]. In [117], six male voices are investigated by applying Fourier analysis 
to one oscillating period of pressure in the speech channel, while in [105, 
106] , the dependence of the base tone on air pressure in the glottic chink is 
determined, and in [24], sound formation and the determination of formants by 
anatomical measurements of the vocal tract using X-rays and electronic computer 
processing of the data are investigated. 

Speech is created by pulses from the vocal cords, the oscillating fre- 
quency of which determines the pitch of the base tone. These sound pulses have 
a discrete spectrum with a large number of harmonics over a broad frequency 



range. The frequency of the base tone lies primarily between 80 and 350 Hz. 
According to the data of some authors, the amplitudes of the harmonics are 
almost identical over a broad frequency range [17], while other authors 
indicate that the amplitude of the harmonics decreases regularly with 
increasing frequency [64]. The audible signal is produced from the sound of 
the vocal cords as it passes through the resonating cavities of the mouth and 
nose, as a result of which the amplitudes of certain harmonics of the vocal 
cord sound are decreased or completely suppressed, while that of others is 
reinforced, forming resonant peaks (so-called formants), the measurement of 
which has also been the subject of a number of works [65, 66, 165]. 

Each voiced sound corresponds to its own combination of formants. 
According to the data of [3], in Russian speech the vowels oo, oh, ah and ee 
are characterized by a single formant, while the sound eh has two and the 
sound y 1 -- 3. According to the data of [64, 66, 165], good recognition of 
vowels requires that the first three formants be determined; according to 
[129], only two need be determined. Unvoiced consonants have no clearly 
expressed formant areas and are distinguished by the amplitudes of the zero 
(M ) , first (M..) and second (M») orders, characterizing the spectrum in the 

band selected [22] : 



M~ = 2A , M, = Sf A , M„ = Sf 2 A 2 , 
n' 1 n n' 2 n n' 



where A is the amplitude of the n-th band of the spectrum, f is its mean 
frequency. 

The sound of the vocal cords h(t) has the frequency spectrum of the even 
and odd portions of the resonators, i.e. 

m 

— » 

The resonators of the oral and nasal cavity, depending on the speech sound 

being formed, have transfer functions G(jw), i.e. G (jw) , G (jw) , G (jw) , etc., 

3- o y 

corresponding to the vowels a, o, y, etc. The signal, upon leaving the mouth, 

is determined by the equation 

m 

— SB 

1 A vowel sound with no exact equivalent in English -- Tr. 



Since the sound of the vocal cords has a spectrum almost identical for / 380 
all people, the audible signal is determined only by the transfer function of 
the resonator (C(jco) = G(jw)H(jw). Each phoneme corresponds to its own 
standard transfer function. On the basis of this phenomenon, a device has been 
developed for recognizing a man from his voice [188], as well as for producing 
information concerning the condition of an astronaut during flight [1]. 

Each sound has its own shades. The timbre of the voice of any man 
differs depending on the properties of the resonators and the change in the 
base tone during speech. The amplitudes and frequencies of the formants of 
each phoneme may change within certain limits [87, 113, 131]. This fact makes 
more difficult the design of a device for automatic recognition and synthesis, 
since the same concentrations of energies at the same frequency may belong to 
different phonemes [110]. 

There are various opinions concerning the significance of the formants. 
Some investigators believe that their definition can have no decisive signif- 
icance in developing a device for recognizing speech signals [79, 121]; 
however, most authors still hold the opposite opinion [24, 67]. For example, 
if we record the vowel ah on magnetic tape and suppress certain formant areas 
within it, the ah is converted, for example, to an oo. 

As we know, most phonemes, including the vowels, can be reproduced by a 
whisper, without using the vocal cords. The exciting action is the breathing 
noise, which can be looked upon as a random, stationary process n(t), with 
spectral density N(w). Then, the spectral density of the output quantity, i.e. 
the phoneme produced, for example ah, will be 

A-,(.,j) = |C a (/iJ)t-A/(o)). 



As we can see from this formula, the base tone carries no information 
concerning the phoneme, and therefore in developing an apparatus for automatic 
speech recognition, there is no need to take the base tone into consideration. 

The transfer function of a phoneme G(jw) contains both amplitude and phase 
data. As we know, the hearing does not react to a change in the phase shifts 
of a complex signal; therefore, in developing an apparatus for automatic 
recognition, these shifts can be ignored; however, in synthesis, a phase shift 
between harmonics worsens the quality of speech sound considerably [35, 41, 108, 
147]. 

Dynamic spectrographs (so-called videographs) have been developed for the 
analysis of speech signals; these videographs make it possible to produce a 
three-dimensional representation of speech: the abscissa represents time, the 
ordinate represents frequency and the third coordinate, the darkness of a point 
on the frequency- time plane, represents the amplitude of the signal at the 
given frequency [128, 137, illeg.] (Figure 2). 












* r -t 




*? 


2 




• 






i 






4*i»- 
















3Mo- 

f»5- 




* - 


i 




f*' 


,*■- 




A-4 

0j- 


»■*■»"" 


' ■ % 


' s. 












tf 




Y .'■? / 


i 


; 





Figure 2. Videograph of Syllables (example taken from [6*t]) 

Videographs are divided according to their principle of operation into / 581 
devices with sequential analysis and devices with parallel analysis. One of 
the first videographs was developed by H. Sund [166]. This device also makes 
it possible to photograph the amplitude-frequency dependence on a cathode ray 
tube. One defect of Sund's videograph is the impossibility of observing and 
recording the time envelopes of the speech signal. 

One modification of the videograph is the intervalograph [44] . In this 
device, a signal is produced whose amplitude is proportional to the intervals 
between transitions of the speech signal through the zero level. 

On request by the Institute of Language and Literature of the Academy of 
Sciences ESSR, the former Scientific Research Electronic Engineering Institute 
of the ESSR Council of the National Economy prepared a spectrograph with 
parallel analysis. A set of 52 filters is used to cover the range from 40 Hz 
to 14 KHz. The spectrogram of the speech signal being analyzed is produced 
simultaneously on three cathode ray tubes, making it possible to photograph 
and visually observe changes in the spectrum of the speech signal. 

A new variant of a spectrograph manufactured by General Electric has 
80 filters with a spectral width of 75 Hz each, a local oscillator, a scanning 
device, an amplitude logarithmizing device, and a camera for photographing the 
results of analysis. The spectrograph can also be used for analysis of the 
songs of birds, machine noises, the operation of the heart, etc. [186, 187]. 

Using the videograph it has been established that the accuracy of recog- 
nition is increased not only by determining the position of formants and their 
intensity, but also by establishing the rate with which they change in 
frequency and level. 

Speech is a continuous function of time between pauses for breathing, and 
consists of individual, discrete phonetic elements -- phonemes. The phonemes 
are varieties of sounds which depend on their pronunciation. There are always 
more phonemes than sounds. In the Russian and English languages there are 
approximately 40, in Estonian there are approximately 30 and in German there 



are approximately 40 [47] . 

The problem of recognition of speech sounds can be solved either by 
comparison of one phoneme, syllable or word from the corresponding set stored 
in the memory of the machine, or by acoustical characteristics alone. In the 
first case, the machine performs comparison of the characteristics of the input 
speech signal with the characteristics of signals stored in its machine memory, 
and the input signal is output as the sound with the greatest correlation 
coefficient. In the second case, the output signal is formed by analysis 
alone. 

The perception of speech by man can be divided into three stages. In the 
first stage, the acoustical stage, a number of physical phenomena are perceived 
and a combination of parameters is determined; in the second stage, the 
phonetic stage, these parameters are compared with standard parameters in memory 
and the initial recognition of the speech sound is performed; finally, 
in the third, linguistic stage, the content of the information produced is 
clarified (for example, endings are added to words which the speaker may not 
have pronounced at all, etc.). 

Automatic recognition of speech sounds is based primarily on the perform- 
ance of the first two stages. In the general case, the machine must contain 
devices storing information concerning the language in order for complete 
recognition of speech sounds to be possible. 

The most widespread methods of recognition of speech sounds are analysis 
of the spectral, time and spectral-time characteristics. 

The spectral method was first described by L. Myasnikov [16, 17, 18], then 
later by others as well [61, 129]. According to his suggestion, the audible 
oscillations are analyzed by pairs of filters with the following pass-bands: 
500-700 and 800-1000 Hz; 1250-1500 and 4000-5000 Hz; 650-750 and 5500-6500 Hz; 
1250-1500 and 400-500 Hz. The output of each filter pair is detected and fed 
in counterphase to an indicator whose arrow either remains at the central 
position or is deflected to one side or the other. The phonemes are differ- 
entiated only according to the frequency, regardless of the amplitude. The 
replacement of the indicator with the arrow by a three-position, polarized 
relay was used to produce a recognition accuracy of up to 75-80%. 

C. Smith used 32 filters in his device [160, 161], and employed the / 382 
principle of accentuation of formant peaks by determining the amplitude 
differences in the adjacent filters. The signals produced were fed to a 
comparison system which reacted only to signals with maximum amplitude. A 
certain combination of channel numbers, i.e. formant frequencies, corresponds 
to a certain vowel phoneme. However, the results were unsatisfactory due to 
the fact that the uncertainties resulting from the influence of neighboring 
phonemes on each other and displacement of formants resulting from changes in 
the base tone could not be eliminated. 



3. Vocoders 

The investigation of formants and the spectral method of analysis were 
used as the basis for construction of band, formant, scanning, harmonic and 
correlation vocoders. 

In the band type vocoder, the entire range of the speech signal is 
analyzed in band filters having either even frequency division throughout the 
entire range or even frequency division only in the lower range (up to 
1000 Hz) with logarithmic frequency division in the higher range (above 
1000 Hz) [104]. The outputs from each filter are rectified and used as para- 
meters for analysis of the speech signal. 

In the first vocoder [59] , the speech signal was analyzed by ten main and 
two supplementary filters. The band width of the first filter was 250 Hz, of 
the others -- 300 Hz; the entire range analyzed was 2950 Hz wide. The output 
from each filter was rectified, passed through a supplementary low frequency 
filter (tuned to 0-250 Hz) and transmitted through a communications channel to 
the synthesizer. The synthesizer was a noise generator with a frequency limit 
of 250-3500 Hz. The output of the generator was connected to the generator 
filter, which had the same frequency ranges as the analyzer. Thus, the output 
of the generator filter was controlled by corresponding output of the analyzer 
filter signal. The base tone was separated from the signal before it passed 
through the analyzer filters and passed through a second, supplementary filter 
with a frequency pass-band of 0-50 Hz for smoothing. 

Since in this vocoder all the higher harmonics of the speech signal are 
not transmitted, the synthesized speech sounds rigid and hoarse, but the 
intelligibility is satisfactory. 

The communications channel carries ten main signals and two supplementary 
signals. Each channel requires a band width of approximately 25 Hz, i.e. a 
total of approximately 300 Hz. 

If we compare uncompressed speech signals with identical volumes of 
information but with different ratios of frequency and dynamic ranges, we have 

whereP , P , P and P are the power of signal and noise at the trans- 

c l n l c 2 n 2 
mitting and receiving ends of the communications channel; F and F„ are the 

corresponding frequency bands; c, and c~ are the throughput capacities for 

transmission of quantities of information. 

Assuming that c. = c ? , i.e. the throughput capacities of ordinary and 
vocoder transmission are identical, where F, = 3000 Hz and F 2 = 300 Hz we have 



P /P = (P /P ) , i.e. theoretically the vocoders require ten times 

c 2 n 2 c l 1 
greater dynamic range. Actually, this requirement is overstated and, according 
to the data of a number of authors [35, 63], if transmission of uncompressed 
speech requires a signal/noise ratio of about 30 db, a band type vocoder with 
ten channels requires a ratio of about 40 db [161]. 

Vocoder signals can be transmitted either continuously or in the form of 
pulses. The continuous signal has a spectral width of not over 15-50 Hz. The 
dynamic range is somewhat less than that of the original speech, not exceeding 
25-30 db. With pulse transmission, vocoder signals are quantized by level and 
time. In frequency, not over two samples per Hertz are required, while in 
amplitude, samples each 1-1.5 db will suffice. With a spectral width of 
25 Hz and a dynamic range of about 16 db, the channel should provide a 
capacity of 200 bits per second (25 X 2 = 50 samples; 2 4 = 16 and 4 X 50 = 200) 
or a total throughput capacity of about 2000 bits per second for a ten-channel 
vocoder. 

We know that the transmission of pulses requires a frequency of no less /383 
than 1-1.5 Hz per bit/sec, so that transmission at a rate of 2000 bit/sec 
requires a channel with a frequency width of 3000 Hz. 

According to information theory, speech can be transmitted without distor- 
tion when the following condition is fulfilled: 

A>^rD F bit/sec 
* av 

where A is the throughput capacity of the channel; D is the average 

clV 

effective dynamic range, F is the width of the frequency range of speech. If 
we consider that the speech signal range is 5000 Hz, the mean dynamic range is 

3 
30 db, then A > 5*10 (bit/sec); consequently, according to the data above, the 

loading of communications channels can be decreased by a factor of approx- 
imately 25. 

In a semi-vocoder, the lower frequency of speech (up to 600 Hz) is trans- 
mitted without conversion, while the high frequency area (above 600 Hz) is 
analyzed and transmitted by a vocoder [69, 155]. In comparison to vocoders, 
the required channel volume in this case is increased, but the intelligibility 
of one-syllable words is increased from 74 to 84%. 

In a scanning vocoder [175, 176, 178], the speech signal is analyzed in 
100 filters, the outputs of which are detected and stored in condensers. The 
voltage level of the condensers is transmitted by a rotating commutator at 
30 rotations per second to a synthesizer, where the sound is restored using a 
controlled multivibrator. Vocoders have also been developed in which the base 
tone is not transmitted [34]; in this case, the intelligibility produced is up 
to 62.5%. 



transmitted in linear combinations. The difference is in the synthesis at the 
receiving end: in band type vocoders, the parameters are fixed by filters with 
the same bands as the analyzers, while harmonic vocoders have filters at the 
receiving end corresponding to the expansion into the Fourier series. 



speech 
input 



separate 

main tone 



formant 
analyzer 



formant 
analyzer 



formant 
analyzer 



tone 
generato r 



modulator 



1 



modu lator 



T 



modulator 



J 



transmission 
channel 



r 



noise 
generator 



tuning of 
frequency 



J 



tune 

frequency 

ci rcui t 



tune 
frequency 
ci rcu i t 



output of 
synthe- 
sized 
speech 



J 



Figure 3- Block Diagram of Formant Vocoder 



The correlation method of analysis is based on the relationships between 
the autocorrelation function R(r) and the energy spectrum of the signal S(o>) 
[23, 33, 154]: 



The correlation method allows us to avoid the influence of the effects of 
phased shifts in the synthesized speech, which appear in band type, formant 



10 



In a pulse vocoder [175], there are ten evenly distributed filters, the 
outputs of which are rectified and used to control pulse generators so that 
pulse width modulation is produced, the number of pulses for each formant being 
proportional in amplitude and frequency. According to the statement of the 
author of [175], this vocoder has great interference stability. 

The spectral-time method of speech recognition differs from the spectral 
method in that the output of the filters is scanned with respect to time as 
well as frequency. As a result, phonemes are transmitted through the communi- 
cations channel in the form of code symbols [140, 142] . 

A further development of the spectral method is the formant method. It 
consists of determination of the presence of a given formant in a given band 
of filters [67-69, 75] (Figure 3). In vocoders of this type, up to four 
formants are distinguished. Improved formant vocoders, in which the intensity 
as well as correlation between frequencies and amplitudes of formants are 
analyzed, have been developed by several authors [3, 50, 64]. Formant 
vocoders in which the moments of first and second order are considered have 
given better results, particularly for determination and synthesis of conson- 
ants than the vocoders described above [43, 63, 85]. It has been suggested 
that the first formant be determined both in the 250-850 Hz range and in the 
300-1200 Hz range. According to [85], it is sufficient to have an upper limit 
to the filter defining the first formant of not over 1000 Hz. The second 
formant is determined between 900 and 2300 Hz. Since the frequencies of 
formants overlap in this vocoder, it has been suggested that the position of 
the first formant be determined first. If it is between 700 and 900 Hz, the 
frequency of the second formant must be no lower than 1400-1800 Hz; if, 
however, the first formant is considerably lower than 700 Hz, the frequency of 
the second formant lies between 700 and 900 Hz and it is necessary to retune 
the filters accordingly. If a third formant is transmitted as well, it will be 
defined in the 2100-3500 Hz band, the fourth -- in the 4000-6000 Hz range. / 384 
The moments are determined by differentiation of the spectrum. If the base 
tone is transmitted by two parameters (frequency and level) and the four 
formants are transmitted by three parameters, 14 signals must be transmitted in 
all. If the moments as well as the dynamic indicators of the formants such as 
rate and direction of change of frequency and level are transmitted, the 
number of signals transmitted is increased. 

Scanning vocoders also have analyzing filters, but the signals are trans- 
mitted sequentially in time to the expander [177] . Whereas in formant 
vocoders the speech signal is analyzed, in contrast to band vocoders, only at 
frequencies corresponding to the position of the formants in the speech signal, 
harmonic vocoders analyze the speech signal completely, determining the Fourier 
coefficients and transmitting the terms of the series (except for the constant 
component) through the communications channel using two parameters. At the 
receiving end, these parameters control either a discrete spectrum generator, 
or a noise generator, sometimes both[20, 21, 27]. In their principle of oper- 
ation, harmonic vocoders differ little from band type vocoders. In both cases, 
the ordinates of the spectrum are determined, except that in band vocoders they 
are transmitted without conversion, while in harmonic vocoders they are 

11 



and harmonic vocoders due to the presence of a complex impedance in the band 
filters of the synthesizer. This achieves compression by a factor of 10 
[153, 154]. 

In order to improve intelligibility of the phonemes in the spectral-time 
method, it has been suggested that all phonemes be preliminarily divided 
according to their characteristic indicators [94, 95]. An electronic binary 
system [43, 45, 185] separates voiced and unvoiced sounds using filters -- the 
voiced sounds have a base tone, the unvoiced sounds do not. In the next step, 
the noise voiced sounds are distinguished from the non-noise voiced sounds 
by the presence or absence of the first formant at the output of the filters. 
Unvoiced sounds are divided into plosive and fricative by the difference in 
their amplitudes. The unit for separating voiced sounds into noise and non- 
noise types operates with an accuracy of up to 95%. The unit for separating 
vowels into higher and lower vowels operates with an accuracy up to 98%, while 
the unit separating lower vowels into diffuse and compact operates with an 
accuracy of up to 94%. 

Comparative data on the volume of the communications channels are shown 
in Table 1 [159]. 



/385 



TABLE 1 





Coding method 


Required channel volume, 


bi t/sec 


discrete form of speech signal 

phoneme 

word (120 words per minute) 

a) vocabulary of 2 words 

b) vocabulary of 8,000 words 
vocoder 

teletype (120 words per minute) 


30.000 1 
60 

2 

26 

2,000 

75 





Considering that the speech signal range is 3,000 Hz. 



A number of works have been dedicated to improvement of the technical 
indicators of vocoders [49, 114, 164]. For example, digital vocoders have been 
developed where in place of exciting the synthesizer with the voice, as was 
done in the first vocoders, the excitation is performed by the base tone 
which in turn is used to turn on a special generator. The synthesizer uses a 
reducing device, which eliminates the noise arising as a result of variation 
of the base tone [171]. A new system of compression has been developed which 
does not require separation of the base tone [76] . The usage of a multi- 
channel modulator has reduced the transmitted signal spectrum to 1.4 KHz [77]. 
A vocoder system has been suggested allowing a considerable reduction in the 
volume of the communications channel (to 94 bit/sec) ; the operation of the 



12 



vocoder is based on the usage of a memory device in which all phonemes are 
stored, each phoneme having its own code and being "called up" by digital 
signals [47] . 

Another suggestion for reducing the volume of the communications channel 
is based on the usage of syllable synthesizers. A code for the syllables in 
the form of digits is obtained in the analyzer, and the digits are then trans- 
mitted to the synthesizer and the required syllable is selected on the basis of 
these code digits. A synthesizer with a volume of 200 syllables has been 
constructed on this basis [126] . 

The EVA electron-analog synthesizer reproduces speech sounds using curves 
drawn with current conducting ink on a tape or drum [96] . The transmission of 
speech by this method can be performed at a frequency 30 times lower than that 
required for ordinary telephone conversation. The intelligibility of words 
reaches 75%. According to a statement by the author of [96], intelligibility 
can be increased to 85%, and the frequency compression increased to 1,000. 
This type of transmission requires very little power, which is extremely 
important in establishing communications with astronauts. 

4. Special Devices for Recognition of Speech Signals 

The spectral-time method was studied in detail by H. Olson and H. Belar 
[122, 123] and J. Dreyfus-Graf [57, 58], who developed typewriters controlled 
by oral commands. The first machine of H. Olson and H. Belar typed words from / 386 
a vocabulary of ten single- syllable words in the memory of the machine; the 
third typewriter had a vocabulary of 150 single-syllable words (Figure 4) . 
In the opinion of the authors, a typewriter with a memory of 1,000 sound 
combinations is sufficient for practical uses [124] (obviously, this require- 
ment is also sufficient for the Russian and Estonian languages, since in these 
languages phonetic transcription and printed form differ little from each 
other) . The latest model of their voice powered typewriter consists of eight 
band-pass filters, eight amplitude- comparing detectors, syllable and ortho- 
graphic memory units, control devices and the typewriter. The filters cover 
the range from 250 to 20,000 Hz. The output of each channel is transmitted to 
a device where the maxima are accentuated by comparing the levels of the 
outputs of neighboring filters and the second derivative of the curve of the 
speech signal is defined. Quantization of the input signal with respect to 
time occurs each 0.2 sec, and after its output from the filters it is 
quantized in the memory each 0.04 sec (in all, also 0.2 sec) and is quantized / 387 
by amplitude in three levels. 

The comparison circuit compares the output curve of the speech signal 
before and after quantization, i.e. each 0.04 sec. If no change in signal 
amplitude occurs during this time, the signal does not reach the memory. The 
memory has 256 different characters, including pause. The results of analysis 
are transmitted in an eight digit code, which at a rate of pronunciation of 
ten phonemes per second requires a rate of 80 bits per second. The accuracy 
of operation of the machine is 92-94% if it is adjusted to the speaker using 

13 



it, considerably lower if random speakers are used. 

oand- 
pass Comparison 



Compressor l^. l4 . «. ., 
K . filters apparatus 



noise 







suppressor £ 
J S 



Expe r i men 
*4icrophon 



ter'fe 
one ! 



in 



tSiL338!}-C 







— 




i_ 


> 


4-> 


L. 


o 


o 


<u 


y 


Q. 


<u 


00 


z 



in 






: Syllable «= 



c c 



oo <u — +i>— : 

^ >. -QO 



sj/elopment <2I I 



la oJ in i-—' 
_. C (U+->0 

o Time ■' (o t 

- — n — *i~ 



~T 



Syllable 
memory 



1-24 



i-^ hf-6| 



25-^1 ^-6^65-9^ 



i9 

— Q 

r- (U 

— > 

oo -a 



Orthographic memory 



1-32 



33-64 



65-95 



> 



C 4-> 

— c 
^- o 

3. O 



£\ 



Typewri ter 



j 



Figure k. Block Diagram of Phonetic Typewriter of H. Olson and 

H. Belar 



H. Olson and his colleagues developed a device in which recognition, 
recoding, printing and synthesis of speech is performed [125]. The memory of 
the machine consists of four English, eight French, four German and four 
Spanish words. The rate of operation of the machine is 60 words per minute. 
The accuracy of the operation with known speaker is 96-98%. 

In the fourth model of the phonetic typewriter, still under development 
by J. Dreyfus-Graf [58], according to an announcement by the author, the 
specific features of the speech of various speakers have been eliminated. As 
in preceding models by this author, there is no memory, so that its accuracy 
of operation depends to a considerable extent on the carefulness of pronunci- 
ation of the speaker. Analysis of the speech spectrum is performed in ten 
filters (400-3200 Hz); also, there are other filters covering the 150-300 and 
4500-6000 Hz range. Each spectral channel contains a detector and a low 



14 



frequency filter for the determination of subformants (0-30 Hz) and a 
quantizer. The rate of change of the speech signal envelope is also 
determined. As a result, the speech signal is separated into 48 signals, each 
of which is quantized by time each one-fifteenth second and by three levels. 
The total information produced amounts to 1440 bit/sec, and the alphabet of - 
the machine has 30 letters. No data on the accuracy of operation of the 
machine have yet been published. 

The writing machine of M. Kalfaian [100] has a device for reducing the 
speech signal to a single base tone frequency using a generator controlled by 
the base tone of the input signal. They are filters whose outputs are 
compared after rectification with data contained in the memory by using 
electronic relays; the machine prints phonemes in a phonetic alphabet, not 
the ordinary written alphabet. No correction for errors in pronunciation is 
performed. The details of the operation of the machine have not yet been 
published. 

The mechanical speech signal recognition machine developed at London 
University has, in contrast to the preceding machines, a linguistic memory. 
The inventor of the machine [51] states that the results of analysis can be 
used to control a typewriter, but he does not expect particularly great 
success. The recognition machine can distinguish four vowels and nine conson- 
ants. The differentiation is performed according to the maximum voltage from 
two of eighteen filters covering the speech signal range from 160 to 8000 Hz. 
It has been established, for example, that the vowel i corresponds to the 
maximum product of the outputs of the filters at 250 and 3200 Hz, the 
consonant m -- to the maximum product at 200 and 320 Hz. Consonants with 
almost identical spectra are distinguished further either according to length 
or voltage level. The linguistic memory contains information on the 
probability of sequences of two phonemes. The device consists of several 
potentiometers, and the information is input to the matrix of potentiometers 
by the position of their slide contacts. The acoustical device determines the 
form of the spectrum, length and intensity of sound and presence of base tone. 
The linguistic device contains standard combinations of phonemes present in 
speech, and combinations which do not exist in speech are forbidden. Thus, 
recognition occurs in the acoustical device and correction, i.e. improvement 
of intelligibility, occurs in the linguistic device. Experiments with 
200 words have produced an accuracy of recognition of 72% (45% for any 
voice) . 

The correlation method of recognition is also based on multiplication of 
filter outputs. A device developed in the Bell Laboratories is designed for 
recognition of ten numbers pronounced by any subscriber [50] . The speech 
spectrum received is multiplied by a typical phoneme spectrum, after which the 
average value is determined. The maximum value for multiplication corresponds 
to the phoneme which is most similar to that received. The device consists 
of a complex containing ten relays, only one relay giving an output pulse for / 388 
any one digit. The accuracy of recognition is 97-99% with a known speaker, 
50-60% with any speaker. 

15 



In Japan, a device has been developed to recognize ten numbers pronounced 
separately. This process is performed using eight characteristics (number of 
voiced intervals per word, position of first and second formants at the 
beginning of the first voiced interval, position 100 msec after beginning of 
first voiced interval, etc.), with from two to thirteen levels each. The 
voltages corresponding to the levels of these eight characteristics are sent 
to a matrix circuit in which the conditional probabilities of all numbers are 
calculated. The highest conditional probability corresponds to the number 
actually pronounced. In pronunciation of one thousand words by one speaker, 
an accuracy of 99.7% was achieved, while when twenty different speakers (all 
men) pronounced one thousand words, an accuracy of 97.9% was achieved [189]. 

The results of an experimental investigation on the determination of 
various characteristics of numbers pronounced in Russian are presented in 
[30]. 

The time method of recognition is based on the usage of clipped speech. 
This method was presented in the works of I. Licklider [112] and subsequently 
in the works of other scientists [145, 150, 174], including Soviet scientists 
[12]. G. Tsemel' [28] uses clipped speech for differentiation of certain 
consonants. According to his data, the accuracy of recognition of the sounds 
p and t was 95%, the sound k -- 75%. 

A. Rais [145] analyzed consonants based on their good differentiability 
by the ear. The experiments were performed to determine the capabilities for 
input of data in oral form to an automatic translation machine. It was 
established that the dependence of the number of pulses of clipped speech on 
the time for various vowels pronounced by different speakers is almost linear. 
The results of the experiments with three speakers are presented in Table 2. 
Preliminary differentiation of the signal was not performed. 





TABLE 2 




Speaker 


a 


o u 


i 




Pu Ises/sec 


first 

second 

third 


558 
533 
491 


450 350 
417 317 
384 292 


283 
218 



I. Toffler suggested a method of separating the base tone from speech 
signals by using nonlinear elements and the clipping method [172]. 

Clipped speech was also used at Kyoto University [150]. Analysis was 
performed both for vowels and for consonants. According to the methodology 
used in this work, the speech signal was amplified to a preselected level, 
then fed to the input of a stage of Kipp oscillators, the outputs of which are 



16 



pulses corresponding to the moments of changes in the direction of the 

rectangular clipped speech signal. The Kipp oscillator stage, the pulse 

lengths from which have different values in different cases, classify the 

clipped speech signal into fourteen values from 0.59-1.11 to 22.55- 

-4 
32.22*10 sec. In each channel, the number of pulses is determined by a 

counter. Analysis of vowels is performed according to the distribution of 

time intervals W(t) and the univariate distribution of probability W 1 (t) 

according to the formulas 

lT(t m/ )= i -g- and W x (x ) = i p^- (i = i... M). 



where t . is the mean time interval in the i-th channel: v . is the number of 

mi ' i 

intervals in the i-th channel; At . is the time difference between the upper 
and lower boundaries of the i-th channel. 

A writing machine based on analysis of clipped speech was constructed at / 389 
the same university [149], and has a memory volume of 200 syllables. The 
details of operation of the machine have not yet been published. 

A device for automatic distinction between twenty spoken words (the 
numbers from zero to nine, plus, minus, space, forward, back, etc.), based on 
clipping of preliminarily differentiated speech signals, was developed at the 
Academy of Sciences of the Georgian SSR [13] . It has been stated that the 
device is capable of distinguishing up to twenty spoken commands when tuned 
for a particular voice, or on the order of five commands from many voices. 

The portable electronic calculating machine SHOEBOX, as yet the only 
series produced model, reacts to spoken pronunciation of the ten numbers (from 
zero to nine) and six additional commands: plus, minus, sum, partial sum, 
error, clear. The word "error" stops the device and clears all operations 
performed. Recognition occurs according to the location of the first phoneme, 
stressed vowel, last phoneme and time envelope of positive and negative peaks 
of the curve of the word. The operation of the device is unstable with 
different speakers [115] . 

Of the three writing machines which we have described [58, 123, 149], 
the most promising device, according to the materials of the Stockholm Seminar 
of 1962 [139] is the machine of J. Dreyfus-Graf, which has no limiting memory 
unit [32, 141]. 

Let us present certain other data concerning special devices for the 
analysis, recognition and usage of speech signals. 

The analysis and recognition of audiofrequency signals can be performed 
using a device which has been developed, the operation of which is based on a 

17 



resonant mechanical system consisting of glass fibers of various lengths and 
diameters up to 0.05 mm. One end of each fiber is fastened to a special 
shaped base, the other end can oscillate freely under the excitation of the 
audio waves. Using light beams, the oscillations of these free ends are 
projected through a standard screen onto photo diodes. The element consists 
of 2,000 fibers with a total volume of 16 cm , analyzing the frequency range 
from 30 to 20,000 Hz with a resolving capacity of about 10 Hz. Each standard 
screen stores the signals of prototypes. The photo element integrates the 
entire quantity of light, and the more closely the signal being analyzed 
corresponds to the standard signal represented by the standard screen, the 
greater the total light flux. If the flux exceeds a given threshold, the 
signal is recognized. Since the standard can also be changed in correspond- 
ence to the information produced, this element has characteristics of self- 
teaching. The device recognizes only short words pronounced by a given voice. 
This device was constructed in an attempt to establish communications with 
dolphins. It was established that the "speech" of dolphins consists of short 
signals (approximately 0.1 sec) at frequencies from 5 to 10 KHz [180, 181]. 

Harmonic analysis of periodic functions fixed in the form of graphs or 
tables can be performed using an electromechanical harmonic analyzer allowing 
simultaneous production of five pairs of coefficients of Fourier series with an 
accuracy to 0.3% of the maximum value of the function being analyzed [4]. 

A device has been constructed in which the recognition of commands 
(numbers) is achieved using visual data on the movement of the lips. Light 
sources with directed reflectors are installed on each lip. A photo resistor 
with a collecting reflector is placed before the lips, connected to one arm 
of a balanced bridge. The voltage taken from the bridge is amplified, fed to 
a differential amplifier and thence to a strip chart recorder. The light 
sources and photo resistors with reflector are connected to the head of the 
speaker. The intelligibility of ten numbers for a concrete speaker was 91%, 
for two different speakers -- 78.3%. By determining one more parameter -- the 
rate of movement of the air flow near the lips -- the intelligibility for a 
single speaker could be increased to 100%, for two speakers -- to 81% [84]. 

A system of solenoids has been theoretically developed, using which 
different codes can be produced for 24,000 English words up to 16 letters 
long; however, this system for word recognition has not yet been constructed. 
Essentially, the system is a memory device in which the curves of speech 
signals are stored in digital form and which can be used to produce momentary 
values of the correlation function of the speech signal [38, 134]. 

Based on the usage of relay systems, a device has been created for / 390 
conversion of digital information to speech information. The sounds of the 
numbers zero to nine and the words "volt," "second," "power" are recorded on 
the tracks of a magnetic drum. The address of each drum track is output by a 
ring counter, and the recording thus selected is sent to an audio amplifier 
[144]. 



18 



The neuron network and auditory apparatus of man have been modeled by- 
several authors [19, 74, 83, 90]; also, a correlation theory of hearing has 
been developed [101]. On the basis of this model, a device has been developed 
for the recognition of sounds and individual words. The device consists of 
an output amplifier, a system of filters and a unit of logic and computer 
circuits modeling the outer and middle ear, the inner ear and the nerve net- 
works respectively. The voiced consonants b, d, g and unvoiced consonants p, 
t and k are differentiated using differentiating logic circuits; however, no 
complete electronic analog has yet been produced [116] . 

Experiments have been performed in the "understanding" of speech which is 
not heard. Information on a phoneme is transmitted through the hand using 
24 vibrators connected to the outputs of a functional model of the auditory 
organ. The numbers from one to nine were recorded on magnetic tape. The 
experimental subject is told if he makes an incorrect answer, and the number 
is repeated. After two to three hours of training, 85% correct answers were 
produced. This experiment is of great significance for the deaf [190]. 

The human voice is not symmetrical about a transitional axis as is, for 
example, the noise spectrum. This peculiarity, "asymmetry of the envelope," 
was used for the creation of a safety switch which turns off a powerful 
machine, such as a machine tool, if the operator shouts [163]. 

A speech signal analyzer has been developed, consisting of 54 gaussian 
type filters [80]. Up to 1000 Hz, the filters have a width of 70 Hz each, 
while over 1000 Hz the width increases by 6.5% each step. The output of each 
filter is rectified and quantized; the quantized currents can be tracked 
visually and, after the proper processing, can be input to an electronic 
computer [169]. 

A speech signal analyzer has been developed which consists of 96 filters 
and covers the frequency range from 30 to 8,000 Hz. The signals are normalized, 
detected, up to the fourth derivative is determined, and the output of the 
analyzer is connected to a triple beam oscilloscope. The process of recog- 
nition is currently being automated [53]. 

The solution of the problem of separating the base tone is of great 
significance both for linguists and for the manufacturer of vocoders and 
for solution of the problem of recognition of speech signals in general [78] . 
Several variants of a special apparatus [42, 56, 147] and devices have been 
developed which have units for determining the autocorrelation function of 
signals in order to separate peak maxima [72], the widths of formants [62, 
170] and other supplementary devices [143, 184]. For example, in [184] it is 
suggested that the phase of the outputs of N filters be changed by 90 degrees. 
These outputs are looked upon as multidimensional vectors which change with 
time. If there is periodicity in the signal, the vector will form a closed 
curve. Five filters of 120 Hz width each are used, covering the frequency 
range from 300 to 900 Hz. 



19 



In order to separate the base tone in a speech signal, a device has been 
developed consisting of elements with nonlinear characteristics and slight 
time constants [151]. Another system for analysis of the frequencies of 
formants and base tone operates on the principle of the tracking filter with 
preliminary transfer of the spectrum of frequencies being analyzed [32] or 
generation of a signal with manual adjustment of the signal to correspond to 
the signal being analyzed [179]. 

5- Universal Computers As Means of Investigating and Recognizing Speech / 391 
Signa Is 

The usage of universal computers for the investigation of speech signals 
began shortly after computers were invented [48, 70, 83] and has expanded each 
year. They are used in connection with spectral analysis of speech signals 
[5, 37, 46, 99], for recognition and synthesis of a selected phoneme combin- 
ation [38, 88, 91, 116], or numbers [52, 71, 152], to separate the base tone 
[46, 72, 81, 120] and determine the parameters of formants [127, 167, 182], 
to investigate the operation of vocoders [132], to determine the possibilities 
for input of speech signals into computers for mathematical modeling of 
communications systems [15, 46], etc. In [120], a method is presented for 
expansion of a speech signal into a Fourier series and determination of its 
constants. The amplitudes of each sequence of frequencies are logarithmized 
and analyzed in a second spectral analyzer. The output of this analyzer is 
the logarithm of the power spectrum and has peak values in the case of 
analysis of voiced phonemes, and has no peak values if the phonemes are 
unvoiced or have no base tone. Since the time of changes of the frequencies 
to the speech signals cause periodic pulsations in the amplitude, spectrum, the 
Fourier transform of the spectrum gives the frequency of the pulsations, 
inversely proportional to the frequency of the base tone of the speech. 

Spectral analysis has been performed for a group of preliminarily 
segmented vowel sounds, using a special device which performed digital coding 
of the speech segments being analyzed. These data were input to the computer 
and a stepped synchronous analysis was performed using the Fourier method [37] . 

In [135], a new technique is suggested for measurement of the frequency 
and widths of formants of a speech signal, based on the theory of Fant [24]. 
A form is assigned to the spectral equations: 

" —tBJ 

f (t) - flo + I e ' (at cos 2nFit + b, sin 2.nf,/), 



where N = 3 and 4 (number of formants). 

A method is presented for determination of the numerical values of the 
coefficients a n , a., b., F. (1 < i < N) by computer using the least squares 

20 



method (B. is the width of the i-th formant and F. is its frequency). Based 
on the experimental materials from analysis of three words (bought, bottle and 
beet) pronounced by two persons twice each, it is affirmed that the first two 
formants for o , i and a are found quite reliably. 

The energy spectrum of speech signals can be calculated by methods other 
than expansion into Fourier series, for example by separation of this spectrum 
into polynomials or representation of the spectrum as a Markov process [98] . 
In [5], the results are presented from a correlation-spectral analysis, while 
[10] presents estimates of the error in the spectral analysis method and [7] 
presents an estimate of the frequency of quantization of the speech spectrum 
with correlation and spectral analysis by computer. 

It is stated in [148] that the problem of recognition of various 
patterns, including speech patterns, can be reduced to the problem of finding 
a class in which a given signal belongs if the total number of classes of 
signals in which the signal may be included is known. This problem is solved 
by minimizing a certain risk function, as a result of which the optimal 
rules are found to be used for solving the problem of recognizing output 
signals from an electronic analyzer, a volume of the auditory helix. 

A method has been suggested for representing a model of a formant as an 
n-dimensional vector, each component of which is a single discrete value of 
the formant at a given moment in time [40]. Thus, the dimensions of an 
n-dimensional space are determined by multiplying the discrete values of 
parameters of the formant by the moments of observation. The computer 
develops this data in two steps, called by the inventor of the process 
learning and recognition. The results of analysis of each of ten phonemes /39_2 
(first letters of the alphabet) pronounced by ten speakers, are presented in 
the form of a matrix [40] . 

It has been suggested that a computer program be developed to calculate 
the energy and envelope of the speech signal over a fixed time interval, the 
frequency of transition of the signal through the zero level and the distribu- 
tion of intervals between zeros in clipped speech throughout the entire time 
interval, the autocorrelation and mutual correlation functions and also to 
perform spectral analysis of the speech signal in order to recognize the 
signals [9]. An analog-digital converter with eight digit readout of vinary 
numbers has been developed for this purpose for input of sound information 
into the computer [6] . 

In [53], a speech signal is analyzed in a 30 channel band type analyzer. 
The computer is used to determine the frequencies F.. , F_ and F each 10 msec. 

The sectors of speech during which the formant frequencies do not change are 
not taken into consideration. An experiment for the recognition of ten 
monosyllabic words spoken three times each by three speakers gave 100% correct 
results. Recognition of speakers by their voices was also successful. 

The problem of separating the base tone has also been the subject of many 
works. In addition to special apparatus outlined above, this work is being 

21 



performed by computer, in most cases together with determination of other 
speech parameters. In [81, 182], a program is developed which is used to take 
a speech signal preliminarily processed in a correlation type analyzer through 
an analog-digital converter, introduce it into a computer for determination of 
the parameters of the base tone and other indicators of the speech signal such 
as: frequency and amplitude of the first three formants, instantaneous signal 
power and rate of change of all quantities. The results of separation of the 
base tone are compared to data produced in [109] , in which the determination 
of the base tone and rate of its change during the pronunciation of individual 
words- and syllables was performed generally automatically, while more precise 
determination was made by manual measurement of distances between peaks on 
oscillographs of the speech signals. Good correspondence of the measurement 
results is noted. 

In [167-169], a process is analyzed for separating the frequencies of 
formants and determining the formants in one-syllable words of the Japanese 
language by computer. First, the speech is preliminarily analyzed in a 
filter system covering the frequency range from 200 to 5900 Hz, then the 
speech is coded by an eight bit binary word and sent to a computer magnetic 
tape recorder. In the computer, the formant frequencies are separated from 
the speech, the rate of change of formant frequencies is determined and the 
second order moment is defined near the mean frequency; the phonemes are 
segmented and phoneme classification of the spectrum is performed. It is 
stated that this method allows error free determination of separately 
pronounced phonemes, as well as almost error free determination and recogni- 
tion of sounds and unvoiced consonants. 

In order to automate the exchange of information between a man and a 
warehouse, a system has been created in which audible speech is analyzed 
according to certain energy characteristics, converted into binary electrical 
signals using a device containing frequency filters, analog-digital converters 
and decoders, is coded onto punched cards and transmitted to an electronic 
computer. A subscriber can access to the computer by telephone and 
receive an answer in various languages (French, English, etc.) [107]. A 
similar system (using the IBM-7770 computer) with a memory volume of 60 words 
can answer 750 simultaneous telephone inquiries concerning prices on an 
exchange [8]. 

A unique trend in the investigation of speech signals by computer, 
resulting from the search for methods to reduce the quantity of information 
concerning the sound of speech, has appeared in the usage of the "analysis- 
synthesis" method, suggested by K. Stevens et al. [86, 98, 127]. In [127], 
the initial determination of speech signal parameters is performed by compar- 
ison of the input spectrum with a spectrum formed in the system as a result 
of combination of six curves for the first formants and six for the second 
formants, i.e. 36 standard spectral curves in all. Eight parameters are 
changed in forming the standard spectra: the frequency and width of the bands 
of the first three formants, the frequency of the fourth formant and the 
position of the zero in the source spectrum. It has been stated that this 
method allows sufficiently precise investigation of speech and the production 

22 



of initial data for the development of speech recognition apparatus. 

In another suggestion, the speech signal is subjected directly to / 393 

mathematical analysis in the computer without supplementary apparatus and is 
approximated by 30 orthogonal functions in the form of an exponentially damped 
Fourier series. When the sound is changed, the numerical values of the 
coefficients are changed, while the functions themselves remain unchanged. 
The symbols produced are used for synthesis [55] . 

Similar works are also being conducted in Japan [91, 98] and the Polish 
People's Republic [97]. 

A decrease in the influence of the subjective properties of the speakers 
on the result of machine recognition can be achieved to a certain extent by 
using the principle of self-tuning and self-teaching. If many people are 
talking, as in an ordinary discussion, this principle is, of course, 
ineffective. In [173] , the results are described of a work on the recognition 
of patterns using these methods in combination with "analysis-synthesis" 
methods. Using 20 characteristics, for example, a great quantity of combin- 
ations can be produced, but many of these characteristics are nonessential, 
so that submatrices are formed including only the essential characteristics. 
The degree of importance of each characteristic and combination of character- 
istics is determined during the course of self- teaching, i.e. by synthesis 
of the required "dictionary." At first, the machine generates all char- 
acteristics and, analyzing the relationships between individual cells, 
determines the weight of each one, generating new relationships and determining 
their weights in turn. Recognition is performed by the comparison method. 
The number of proper answers achieved by recognizing known stylized portraits 
and spectrograms of speech was 80-100%, for unknown portraits and spectrograms 
60-100%. 

An algorithm for machine recognition using elements of self-tuning was 
also developed in [183] . With a memory volume of eighteen words, the 
accuracy of differentiation achieved by this self-tuning system in processing 
information input to the machine in four languages was 96%; when twenty words 
were used, the accuracy was 86%. The system could successfully distinguish a 
voice concerning which no preceding experience had been accumulated. 

A program was developed which, with a limited memory volume (numbers from 
zero to nine, plus, minus, equals, parentheses, etc., 83 words in all) the 
computer "learned" to recognize these words both for a known and for a random 
speaker (although with poorer results) , to perform arithmetic operations on 
command with the number introduced by voice, to print out the results and to 
translate them into another language [132]. 

A computer machine for the recognition of speech signals not only has the 
advantage that it makes it possible to use partial information and algorithms 
developed for the machine for translation from one language to another, but 
also has the advantage that its large memory volume and high operating speed 
allow speech signals to be separated for analysis into extremely short 

23 



amplitude and frequency sectors, and allows data to be compared with standard 
data quite rapidly. On the other hand, in order to achieve universality of 
recognition, i.e. independence of the results of analysis from subjective 
properties of the speakers, it should not contain comparison elements; also, 
large computers are expensive and cannot be mobile. Therefore, in spite of 
certain advantages over special machines, the latter, i.e. direct analysis 
machines, are more promising for the final solution of the problem of machine 
recognition of speech signals and usage of these signals in various control 
and communications systems. 

In this article we have touched upon the most general part of the work 
in the area of machine recognition and have not discussed the problem of 
speech synthesis at all. The main factor reducing the accuracy of the oper- 
ation of speech recognition apparatus is the inconstancy of the formant 
indicators, depending to a great extent on the individualities of the speakers, 
and the difficulty of segmenting. However, the results of investigations 
have already been used in communications technology, as well as military 
affairs. For example, vocoders have been developed in the USA which are used 
in aviation [2, 133]. As the apparatus is improved, the area of their 
application will doubtless increase, as a result of which cybernetic systems 
will receive new elements which will react to spoken commands without inter- 
mediate coding. 

REFERENCES 

1. "Analysis of Man's Behavior by His Speech," Elektronika, vol. 37, No. 9, /3JM 

1964. 

2. "The Army Orders a Miniature Vocoder," Elektronika, vol. 37, No. 15, 1964. 

3. Varshavskiy, L. A., I. M. Litvak, "Investigation of Formant Composition and 

Certain Other Physical Characteristics of the Sounds of Russian Speech," 
Problemy Fiziol. Akustiki, No. 3, 1955. 

4. Vasilenko, A. T. , Yu. N. Denisov, "An Electromechanical Harmonics Analyzer," 

Pribory i Tekhnika Eksperimenta , No. 6, 1963. 

5. Voloshin, G. Ya. , "Spectral Analysis of Speech Signals by Electronic 

Computer," Sbornik Trudov Inta Matematiki SO AN SSSR^ Vyahislitel'nyye 
Sistemy [Collected Works of Institute of Mathematics, Siberian Department 
Academy of Sciences USSR, Computer Systems], No. 10, Novosibirsk, 1964. 

6. Voloshin, G. Ya. , "Analog-Digital Converter for Input of Speech Signals to 

Electronic Computers," Sbornik Trudov Inta Matematiki. SO AN SSSE^ 
Vyahislitel'nyye Sistemy [Collected Works of Institute of Mathematics, 
Siberian Department Academy of Sciences USSR, Computer Systems] , No. 10, 
Novosibirsk, 1964. 

7. Voloshin, G. Ya. , "The Frequency of Sampling of a Random Function with 

Correlation-Spectral Analysis," Sbornik Trudov Inta Matematiki SO AN 
SSSEj Vyahislitel'nyye Sistemy [Collected Works of Institute of Mathe- 
matics, Siberian Department Academy of Sciences USSR, Computer Systems], 
No. 14, Novosibirsk, 1964. 

8. "The Computer Answers the Telephone," Elektronika, vol. 37, No. 5, 1964. 

24 



9. Zagoruyko, N. G., "The Exchange of Oral Information Between Man and 

Computer Systems," Sbornik Trudov Inta Matematiki SO AN SSSR^ Vychisli- 
tel'nyye Sistemy [Collected Works of Institute of Mathematics, Siberian 
Department Academy of Sciences USSR, Computer Systems], No. 10, Novo- 
sibirsk, 1964. 

10. Zagoruyko, N. G., "Error in Calculating Energy and Envelope of Speech 

Signal by Computer," Sbornik Trudov Inta Matematiki SO AN SSSR> Vychis- 
litel'nyye Sistemy [Collected Works of Institute of Mathematics, Siberian 
Department Academy of Sciences USSR, Computer Systems], No. 10, Novo- 
sibirsk, 1964. 

11. Zagoruyko, N. G. , G. Ya. Voloshin, V. N. Yelkina, "Automatic Recognition 

of Speech Patterns," Sbornik Trudov Inta Matematiki SO AN SSSR, 
Vyohislitel'nyye Sistemy [Collected Works of Institute of Mathematics, 
Siberian Department Academy of Sciences USSR, Computer Systems], No. 14, 
Novosibirsk, 1964. 

12. Kakauridze, A. G., "Some Problems of Coding of Speech Vowel Sounds," 

Trudy Inta Elektroniki, Avtomatiki i Telemekhaniki AN GruzSSR [Works of 
the Institute of Electronics, Automation and Telemechanics, Academy of 
Sciences Georgian SSR] , No. 1, 1960. 

13. Kakauridze, A. G., "Experimental Device for Automatically Distinguishing 

Among a Limited Set of Speech Commands," Elementy Vyahislitel'noy 
Tekhniki i Mashinnyy Perevod^ In-t. Elektroniki 3 Avtomatiki i Teleme- 
khaniki AN GruzSSR [Elements of Computer Equipment and Machine 
Translation, Institute of Electronics, Automation and Telemechanics, 
Academy of Sciences Georgian SSR], Tbilisi, 1964. 

14. Caldwell, V., E. Glasser, D. Stewart, "An Analog Model of the Ear," 

Sbornik Problemy Bioniki [Collection of Problems of Bionics] , Moscow, 
1965. 

15. Lebman, Yu. A., V. N. Sobolev, "Analog-Digital Converter for Input of 

Speech Signal to Computer Machine," Elektrosvyaz ' , No. 8, 1963. 

16. Myasnikov, L. L. , "Objective Recognition of Speech Sounds," Zh. Tekhnich. 

Fiz. , vol. 13, No. 3, 1943. 

17. Myasnikov, L. L. , "The Sounds of Speech and Their Objective Recognition," 

Vestnik Leningr. Universiteta , No. 3, 1946. 

18. Myasnikov, L. L., "Physical Investigation of the Sounds of Russian Speech," 

Izvestiya Akad. Nauk SSSR, Seriya Fiz. , v. 13, No. 6, 1949. 

19. Myuler, P., T. Martin, F. Puttsrat, "General Principles of Operation in 

Neuron Networks and Their Application to the Recognition of Acoustical 
Patterns," Sbornik Problemy Bioniki, Moscow, 1965. 

20. Pirogov, A. A. , "Theoretical Considerations Concerning a Method for Coding 

and Synthesis of Speech Information Using a Harmonic Function," 
Dokl. na Vsesoyuzn. Soveshoh. Sektsii Rechi Komissii po Akustike AN SSSR 
[Report at Ail-Union Conference of Speech Section of Commission on 
Acoustics, Academy of Sciences USSR], Moscow, 1958. 

21. Pirogov, A. A., "Harmonic System for Compression of Speech Spectra," 

Elektrosvyaz ' , No. 3, 1959. 

22. Sapozhkov, M. A., Reohevoy Signal v Kibernetike i Svyazi [The Speech 

Signal in Cybernetics and Communications], Moscow, 1963. 

23. Solodovnikov, V. V., Statisticheskaya Dinamika Linenykh Sistem Avtomati- 

eheskogo Upravleniya [Statistical Dynamics of Linear Automatic Control 



25 



Systems], Moscow, 1960. 

24. Fant, G., Akusticheskaya Teoriya Recheobrazovaniya [The Acoustical Theory 

of Speech Formation], Moscow, 1964. 

25. Kharkevich, A. A., Spektry i Analiz [Spectra and Analysis], Moscow, 1953. 

26. Kharkevich, A. A. , Ocherki Obshchey Teorii Svyazi [Essays on the General 

Theory of Communications], Moscow, 1955. 

27. Khrapovistkiy, A. V., "Theoretical and Experimental Investigation of a 

Cosine-Logarithmic Vocoder," Dokl. na Vsesoyuzn. Sovesheh. Sektsii 
Reohi Komissii po Akustike AN SSSR [Report at All -Union Conference of 
Speech Section of Commission on Acoustics, Academy of Sciences USSR], 
Moscow, 1959. 

28. Tsemel', G. I., "Determination of the Invariant Characteristics of 

Plosive Sounds on the Basis of Clipped Speech Signals," Izvestiya 

Akad. Nauk SSSR Otdel Tekhnicheskikh Nauk Energetika i Avtomatika, No. 4, 

1959. 

29. Tsemel', G. I., "Automatic Recognition of Speech Sounds," Zavubezhnaya 

Radioelektronika , No. 4, 1962. 

30. Tsemel', G. I., "Recognition of a Small Group of Words by Characteristic 

Features of the Speech Signal," Sbornik Problemy Peredachi Informatsii, 
No. 16, Moscow, 1964. 

31. Chistovich, L. A., "Influence of Frequency Limitations on Intelligibility /395 

of Russian Consonant Sounds," Dokl. na Vsesoyuzn. Sovesheh. Sektsii 
Reohi Komissii po Akustike AN SSSR [Report at All -Union Conference of 
Speech Section of Commission on Acoustics, Academy of Sciences USSR], 
Moscow, 1956. 

32. Abstracts of Papers on Speech Analysis, Stockholm Speech Communications 

Seminar, 1962, Royal Institute on Technology, Stockholm, Sweden, The 
Journal of the Acoustical Society of America (=JASA) , Vol. 35, No. 7, 
1963. 

33. Bennett, W. R. , "The Correlatograph. A Machine for Continuous Display of 

Short Term Correlation." Bell System Techn. J. 3 Vol. 32, Sept. 1953. 

34. Billings, A.R. , "Simple Multiplex Vocoder," Electronic and Radio Engr 3 

Vol. 36, No. 5, 1959. 

35. Billings, A.R. , "Communication Efficiency of Vocoders Comparison of Low- 

Power and Conventional Systems," Electronic and Radio Engr 3 Vol. 36, 
No. 12, 1959. 

36. Bogert, B. P., "Vobanc - a Two-to-One Speech Band-Width Reduction System," 

JASA, Vol. 28, No. 3, 1956. 

37. Borenstein, D.P., "Spectral Characteristics of Digit-Simulating Speech 

Sounds," Bell System Techn, J. 3 Vol. 42, No. 6, 1963. 

38. Brick, D.B., and G. G. Pick, "Microsecond Word-Recognition System," IEEE 

Trans. Electronic Comput. EC- 13, No. 1, 1964. 

39. Campanella, S.A. , "A Survey of Speech Bandwidth Compression Techniques," 

IRE Trans. Audio 3 AU-6, No. 5, 1958. 

40. Campanella, S. I. , Coulter, D. C. , and P. Engler, "Speech Recognition by 

Formant Pattern Matching in N-dimensional Space," JASA, Vol. 36, No. 5, 
1964. 

41. Campanella, S. I., Coulter, D. C. , and R. Irons, "Influence of Transmission 

Error on Formant Coded Compressed Speech Signals," J. Audio Engng. Soc. 3 
Vol. 19, No. 2, 1962. 

26 



42. Carre, R. , Lancia, R. , Paille, J., and R. Gsell, "Etude et realisation 

d'un detecteur de melodie pour analyse de la parole," Onde electr. , 
Vol. 43, No. 434, 1963. 

43. Chang, S. H. , "Two Schemes of Speech Compression System," J AS A, Vol. 28, 

No. 4, 1956. 

44. Chang, S. H. , Pihl, C. E. , and I. Wiren, "The Intervalgram as a Visual 

Representation of Speech Sounds," JASA, Vol. 23, No. 6, 1951. 

45. Cherry, C. , Halle, M. , and R. Jacobson, "Toward a Logical Description of 

Languages in Their Phonemic Aspects," Language, Vol. 29, No. 1, 1953. 

46. Clapper, C. L. , "Digital Circuit Techniques for Speech Analysis," IEEE 

Trans. Communication and Electronics , CE-11, No. 66, 1963. 

47. Cramer, B. , "Sprachsyn these zur Obertragung mit sehr geringer Kanalkapazitat, 

Eaahvichtentechn. Z. ,No. 8, 1964. 

48. David, E. E. , and H. S. McDonald, "Note on Pitch-Synchronous Processing of 

Speech," J AS A, Vol. 28, No. 6, 1956. 

49. David, J., Schroeder, M. K. , Logan, B. F. , and H. J. Prestigiacomo, 

"Voice-Excited Vocoders for Practical Speech Bandwidth Reduction," IRE 
Trans. Inform. Theory, IT-8, No. 5, 1962, 

50. Davis, K. H. , Bidulph, R. , and S. Balashek, "Automatic Recognition of 

Spoken Digits," JASA, Vol. 24, No. 6, 1952. 

51. Denes, P., "The Design and Operation of the Mechanical Speech Recognizer 

at University College London," J. Brit. IRE, Vol. 19, No. 4, 1959. 

52. Denes, P., and M. V. Mathews, "Spoken Digit Recognition Using Time -Frequency 

Pattern Matching," J ASA, Vol. 32, No. 11, 1960. 

53. Deuber, G. , "Varied Approaches Used to Develop System for Reliable 

Recognition of Voice Commands," Electronic News, Vol. 8, No. 383, 1963. 

54. Deweze, A., "Techniques des reconnaissance automatique des formes visuelles 

et sonores," Automatisme , Vol. 9, No. 3, 1964. 

55. Dersch, W. C. , "Speech Operated Safety Switch," Electronics, Vol. 36, No. 

25, 1963. 

56. Dolansky, I. 0., "Instantaneous Pitch-Period Indicator," JASA, Vol. 27, No. 

1, 1955. 

57. Dreyfus-Graf, J., "Phonetographe et Subformants," Bull. Techn. PIT, Bern, 

No. 2, 1957. 

58. Dreyfus-Graf, J., "Phonetographe: Present et Future," Bull. Techn. PIT, Bern, 

No. 5, 1961. 

59. Dudley, H. W. , "Remaking Speech," JASA, Vol. 11, No. 2, 1939. 

60. Dudley, H. W. , "Speech Analysis and Synthesis System," JASA, Vol. 22, No. 

6, 1950. 

61. Dudley, H. W. , "Phonetic Pattern Recognition for Narrowband Transmission," 

JASA, Vol. 30, No. 8, 1958. 

62. Dunn, H. K. , "Methods of Measuring Vowel Formant Bandwidths," JASA, Vol. 33, 

No. 12, 1961. 

63. Fano, R. M. , "The Information Theory Point of View in Speech Communication," 

JASA, Vol. 22, No. 10, 1950. 

64. Fant, G. , "Acoustic Theory of Speech Production with Calculations Based 

on X-Ray Studies of Russian Articulations," Mouton & Co, s'Gravenhage, 
1960. 

65. Fant, G. , Fintoft, K., Liliencrants, J., Lindblom, B. , and J. Martony, /396 

"Formant-Amplitude Measurements," JASA, Vol. 35, No. 11, 1963. 

66. Flanagan, I. L. , "A Difference Limen for Vowel Formant Frequency," JASA, 

Vol. 27, No. 3, 1955. 

27 



67. Flanagan, J. L. , "Automatic Extraction of Formant Frequences from 

Continuous Speech," JASA, Vol. 28, No. 1, 1956. 

68. Flanagan, J. L. , "Band-Width and Channel Capacity Necessary to Transmit 

the Formant Information of Speech," JASA, Vol. 28, No. 4, 1956. 

69. Flanagan, J. L. , "A Resonance-Vocoder and Baseband Complement: A Hybrid 

System for Speech Transmission," IKE. Trans. Audio, AU-8, May-June 1960. 

70. Forgie, J. W. , and C. W. Hughes, "A Real-Time Speech Input System for a 

Digital Computer," JASA, Vol. 30, No. 7, 1958. 

71. Forgie, J. W. and C. D. Forgie, "Results Obtained from a Vowel Recognition 

Computer Programme," JASA, Vol. 31, No. 9, 1959. 

72. Fujisaki, H. , "Automatic Extraction of Fundamental Period of Speech by 

Auto-Correlation Analysis and Peak Detaction," JASA, Vol. 32, No. 11, 
1960. 

73. Gabor, D. , "New Possibilities in Speech Transmission," J. IKE, Vol. 94, 

Nov. 1948. 

74. Galdwell, W. F., "Recognition of Sounds by Cochlear Patterns," IEEE Trans. 

Military Electronics, MIL- 7, No. 2-3, 1963. 

75. Gerstman, L. J., Liberman, A. M. Delattre, P. C. , and F. S. Cooper, "Rate 

and Duration of Change in Formant Frequency as Cues for Identification 
of Speech Sounds," JASA, Vol. 26, No. 9, 1954. 

76. Gold, B. , and C. Rader, "Bandpass Compressor: A New Type of Speech- 

Compression Device," JASA, Vol. 36, No. 6, 1964. 

77. Golden, R. M. , MacLean D. I., and A. I. Prestigiacomo, "Frequency Multiplex 

System for a 10-Spectrum-Channel Voice-Excited Vocoder," JASA, Vol. 36, 
No. 10, 1964. 

78. Gribenski, A., "The Pitch of Sound, Its Measuring and Perception," Nature, 

Vol. 173, No. 4, 1957. 

79. Haggard, M. P., "In Defense of the Formant," Vhonetiea, Vol. 10, No. 3-4, 

1963. 

80. Harris, C. M. , and W. M. Waite, "Gaussian-Filter Spectrum Analyzer," 

Vol. 35, No. 4, 1963. 

81. Harris, C. M. and M. R. Weiss, "Pitch Extraction by Computer Processing of 

High-Resolution Fourier Analysis Data," JASA, Vol. 35, No. 3, 1963. 

82. Hellwarth, G. A., "Speech-Formant Measurement with a Continuously Tuned 

Automatic Tracking Filters," JASA, Vol. 35, No. 5, 1963. 

83. Heydeman, P. , "Ein Modellversuch zum Frequenzunterscheidungsvermogen des 

Ohres," Aoustioa, Vol. 13, No. 2, 1963. 

84. Hillix, W. , "Use of Two Nonacoustic Measures in Computer Recognition of 

Spoken Digits," JASA, Vol. 35, No. 12, 1963. 

85. Howard, C. R. , "Speech Analysis Synthesis Scheme Using Continuous Parameter," 

JASA, Vol. 28, No. 6, 1956. 

86. Howard, C. R. , Chang, S. H. , and M. J. Carrabes, "Analysis and Synthesis of 

Formants and Moments of Speech Spectra," JASA, Vol. 28, No. 4, 1956. 

87. Huggins, W. H. , "A Phase Principle for Complex-Frequency Analysis and its 

Implications in Auditory Theory," JASA, Vol. 24, No. 6, 1952. 

88. Hughes, G. W. , "Identification of Speech Sounds by Means of a Digital 

Computer," JASA, Vol. 31, No. 1, 1959. 

89. Husson, R. , "Zur Spektraistruktur menschlicher Vokale aller Stimmstarken," 

Phonetiea, No. 1-2, 1963. 

90. Inomata, S. , "An Auditory Pattern Processing Model," IEEE Trans. Inform. 

Theory, IT-9, No. 4, 1963. 

28 



91. Inomata, S. , "Speech Recognition and Generation by a Digital Computer," 

Res. Electrotechn. Labs, No. 645, 1963. 

92. Ithell, A. H. , "A Determination of the Acoustical Input Impedance Character- 

istics of Human Ears," Acustica, Vol. 13, No. 4, 1963. 

93. Jaffe, J., Cassotta, L. , and S. Feldstein, "Markovian Model of Time 

Patterns of Speech," Science, Vol. 144, No. 3260, 1964. 

94. Jakobson, R. , Fant, G. G. , and M. Halle, "Preliminaries to Speech Analysis, 

The Distinctive Features and Their Correlates," Massachusetts Institute 
of Technology. Aooust. Lab. Teahn. Rept. No. 13, 1952. 

95. Jakobson, R. , "Die Verteilung der stimmhaften und stimmlosen Gerauschlaufe 

im Russischen, Festschrift fur Max Vasmer," Berlin, 1956. 

96. Johnson, W. , "System to Generate Speech from Written Pattern Shown," 

Electronic News, Vol. 8, No. 392, 1963. 

97. Kacprowski, J., "Speech Compression by Means of Analysis-Synthesis Methods," 

Polish Academy of Sciences, Proa. Vibration Probl. , Vol. 5, No. 3, 1964. 

98. Kadokava, J. and K. Nakata, "Formant Frequency Extraction by Analysis-by- /397 

Synthesis Technique," J. Radio Res. Labs, Vol. 10, No. 49, 1963. 

99. Kadokawa, J., and K. Nakata, "Analysis of Speech by Vocal Tract Configu- 

ration," J. Radio Res. Labs, Vol. 11, No. 54, 1964. 

100. Kaliaian, M. V., "Phonetic Typewriter of Speech," JASA, Vol. 36, No. 6, 

1964. 

101. Karplus, H. B. , "Correlation Hypothesis to Explain the Fine Frequency Dis- 

crimination of the Ear," JASA, Vol. 35, No. 5, 1963. 

102. Klass, P. I., "Vocoder Increases Channels Security," Aviat. Week, Vol. 73, 

No. 4, 1960. 

103. Klumpp, R. G., and I. C. Webster, "Intelligibility of Time-Compressed 

Speech," JASA, Vol. 33, No. 3, 1961. 

104. Loenig, W. , "A New Frequency Scale for Acoustic Measurements," Bell Labs, 

Rec. , Vol. 27, No. 7, 1949. 

105. Ladefoged, P., "Acoustic Correlate of Subglottal Activity," JASA, Vol. 35, 

No. 5, 1963. 

106. Ladefoged, P., and N. P. McKinney, "Loudness, Sound Pressure and Subglottal 

Pressure in Speech," JASA, Vol. 35, No. 4, 1963. 

107. Latil de, P., "La parole est aux calculateurs ," Electronique Industr. 

No. 76, 1964. 

108. Lehiste, I., and G. E. Peterson, "Some Basic Considerations in the Analysis 

of Intonation," JASA, Vol. 33, No. 4, 1961. 

109. Lieberman, P., "Perturbations in Vocal Pitch," JASA, Vol. 33, No. 5, 1961. 

110. Liberman, A. M. , Ingeman, F. , Lisker, L. , Delattre, P. C. , and F. S. 

Cooper, "Minimal Rules for Synthesizing Speech," JASA, Vol. 31, No. 11, 
1959. 

111. Licklider, I. C. , "Effects of Amplitude Distortion Upon the Intelligibility 

of Speech," JASA, Vol. 18, No. 2, 1964. 

112. Licklider, I. C. , "The Intelligibility of Amplitude-Dichotomized. Time- 

Quantized Speech Waves," JASA, Vol. 22, No. 6, 1950. 

113. Licklider, I. C. , "Influence of Phase Coherence Upon the Pitch of Complex 

Periodic Sounds," JASA, Vol. 27, No. 5, 1955. 

114. Licklider, I. C, "Man-Computer Symbiosis," IRE Trans. HFL-1, No. 1, 1960. 

115. "Machines Controlled by Spoken Commands," Datamation, Vol. 8, No. 6, 1962. 

116. Martin, T. B. , and I. I. Talvagel, "Application of Neural Logic to Speech 

Analysis and Recognition," IEEE Trans. Military Electronics , MIL- 7, 
No. 2-3, 1963. 

29 



117. Miller, A. E., and M. V. Mathews, "Investigation of the Glottal Wave- 

shape by Automatic Inverse Filtering," JASA, Vol. 35, No. 11, 1963. 

118. Miller, R. L. , "Improvements in the Vocoder," JASA, Vol. 25, No. 4, 1953. 

119. Nicolau, E., Weber, L. , and S. Gavat, "Aparate pentru recunoasterca 

automata a vocalelor," Automatica ei eleatvonica, Vol. 7, No. 6, 1963. 

120. Noll, A. M. , "Short-Time Spectrum and 'Cepstrum' Techniques for Vocal- 

Pitch Detection," JASA, Vol. 36, No. 2, 1964. 

121. Oeren, F. W. , "Some Critical Observation on the Formant Theory of Vowel 

Recognition," Phonetica, Vol. 10, No. 1-2, 1963. 

122. Olson, H. F. , and H. Belar, "Phonetic Typewriter," JASA, Vol. 28, No. 6, 

1956. 

123. Olson, H. F., and H. Belar, "Phonetic Typewriter III," JASA, Vol. 33, 

No. 11, 1961. 

124. Olson, H. F., and H. Belar, "Syllable Analyzer, Coder and Synthesizer for 

the Transmission of Speech," IRE Trans. Audio, AU-10, No. 11, 1962. 

125. Olson, H. F. , Belar, H. , and R. Sobrino, "Demonstration of a Speech 

Processing System Consisting of a Speech Analyzer, Translator, Typer 
and Synthesizer," JASA, Vol. 34, No. 10, 1962. 

126. Olson, H. F. , and H. Belar, "Performance and a Code-Operated Speech 

Synthesizer," JASA, Vol. 38, No. 5, 1964. 

127. Paul, A. P. House, A. S. and K. N. Stevens, "Automatic Reduction of Vowel 

Spectra: An Analysis-by-Synthesis Method and its Evaluation," JASA, 
Vol. 36, No. 2, 1964. 

128. Peterson, G. E., "Design of Visible Speech Devices," JASA, Vol. 26, No. 3, 

1954. 

129. Peterson, E. , and F. S. Cooper, "Peakpicker: A Band-Width Compression 

Device," JASA, Vol. 29, No. 6, 1957. 

130. Peterson, G. E. and I. Lehiste, "Identification of Filtered Vowels," JASA, 

Vol. 31, No. 6, 1959. 

131. Peterson, G. E. , Sivertsen, E., and D. L. Subrahmanyam, "Intelligibility 

of Diphasic Speech," JASA, Vol. 28, No. 3, 1956. 

132. Petrick, S. R. , "Talking to a Computer," New Scientist, No. 235, 1961. 

133. Phyie, D. L. and I. E. Toffier, "Some Features of the Army Channel 

Vocoder," JASA, Vol. 35, No. 12, 1963. 

134. Pick, G. G.,Gray, C. B., and D. B. Brick, "The Solenoid Array - a New 

Computer Element," IEEE Trans. Electronic Computers, EC- 13, No. 1, 1964. 

135. Pinson, E. N. , "Pitch-Synchronons Time-Domain Estimation of Formant Fre- 

quencies and Bandwidth," JASA, Vol. 35, No. 8, 1963. 

136. Pollack, I., and L. Picket, "Effect of Noise and Filtering on Speech 

Intelligibility at High Levels," JASA, Vol. 29, No. 12, 1957. 

137. Potter, R. K. , Knopp, G.A. , and H. C. Green, "Visible Speech," Van /398 

Nostrand, New York, 1947. 

138. Potter, R. K. and J. C. Steinberg, "Toward the Specification of Speech," 

JASA, Vol. 22, No. 6, 1950. 

139. Proceedings of the Speech Communication Seminar, Stockholm, August 29 

to September 1, 1962, Publ. Speech Transmission Lab. Royal Inst. 
Technology, Stockholm, 1964. 

140. Pruzansky, S. , and P. D. Briener, "Automatic Talker Recognition Using 

Time- Frequency Pattern Matching," JASA, Vol. 33, No. 6, 1961. 

141. Pun, L., "The Phonetograph," Control, Vol. 7, January 1963. 

142. Rader, C. , "Spectra of Vocoder-Channel Signals," JASA, Vol. 35, No. 5, 

1963. 

30 



«£■ 



143. Rader, C. M. , "Vector Pitch Detection," JASA, Vol. 36, No. 10, 1964. 

144. Rawley, I. R. , "Converting Digital Data to Voice," Electronic Industv. 

Vol. 23, No. 4, 1964. 

145. Rais, A., "Vowel Recognition in Clipped Speech," Nature, Vol. 181, No. 3, 

1958. 

146. Righini, G. U. , "A Pitch. Extractor of the Voice," Acustica, Vol. 13, 

No. 4, 1963. 

147. Rosen, G. , "Dynamic Analog Speech Synthesizer," JASA, Vol. 30, No. 3, 

1958. 

148. Sackschewsky, V. E., and H. L. Oestreicher, "Pattern Recognition as a 

Problem in Decision Theory and an Application to Speech Recognition," 
IEEE Trans. Military Electronics, MIL-7, No. 2-3, 1963. 

149. Sakai, T. , Dochita, S., Nagata, K. I., "Phonetic Typewriter," JASA, Vol. 

35, No. 7, 1963. 

150. Sakai, T. , and S. I. Inone, "New Instruments and Methods for Speech 

Analysis," JASA, Vol. 32, No. 4, 1960. 

151. Schief, R. , "Koinzidenz-Filter als Modell fur das menschiche Tonbohenun- 

terscheidungsvermogen," Kybernetik, Vol. 2, H. L, 1963. 

152. Scholtz, P. N. and R. Bakis, "Spoken Digit Recognition Using Vowel- 

consonants Segmentation," JASA, Vol. 34, No. 1, 1962. 

153. Schroder, M. R. , "New Approach to Time Domain Analysis and Synthesis," 

JASA, Vol. 31, No. 6, 1959. 

154. Schroder, M. R. , and T. H. Crystal, "Auto- Correlation Vocoder," JASA, 

Vol. 32, No. 7, 1960. 

155. Schroder, M. R. , and E. E. David, "A Vocoder for Transmitting 10 kc/s 

Speech Over a 3.5 kc/s Channel," Acustica, Vol. 10, No. 1, 1960. 

156. Seki, H. , "A New Method of Speech Transmission by Frequency Demultipli- 

cation and Multiplication," J. Acoust. Soc. Japan, 14 June 1958. 

157. Siedler, G. , "Untersuchungen uber die Bedeutung bestimmter Tonfrequenz- 

bander fur die Verstandlichkeit synthetischer Sprache und uber Anderung 
der Sprachverstandlichkeit bei Kanalvertauschungen," Z. angew. Phys. 
Vol. 13, Nr. 6, 1961. 

158. Simmons, P. L. , "Automation of Speech, Speech Synthesis and Synthetic 

Speech. A Bibliographic Survey from 1950-1960," IRE Trans. Audio, AU-9, 
November- December 1961. 

159. Slaymaker, F. H. , "Bandwidth Compression by Means of Vocoder," IRE Trans. 

Audio, AU-9, January -February 1960. 

160. Smith, C. R. , "A Phoneme Detector," JASA, Vol. 23, No. 4, 1951. 

161. Smith, C. R, , "The Analysis and Automatic Recognition of Speech Sounds," 

Electronic Engng. Vol. 24, No. 8, 1962. 

162. Smith, S. L. , "Man-Computer Information Transfer," Electro-Technology , 

Voo. 72, August 1963. 

163. Speech Recognition Gets Push From Synthesizer, Electronics, Vol. 34, No. 

16, 1961. 

164. Steele, K. W. , and L. E. Cassel, "Quality Improvement in the Channel Vo- 

coder," JASA, Vol. 35, No. 5, 1963. 

165. Stevens, K. N. , "Acoustical Analysis of Speech," JASA, Vol. 30, No. 7, 

1958. 

166. Sund, H. , "A Sound Spectrometer for Speech Analysis," Trans. Royal Inst. 

Technology, Stockholm, No. 112, 1957. 

167. Suzuki, I., Kadokawa, J., Nakata, K. , "Formant- Frequency Extraction by 

the Method of Moment Calculations," JASA, Vol. 35, No. 9, 1963. 



31 



n^m^¥ 



168. Suzuki, I., and K. Nakata, "Phonemic Classification and Recognition of 

Japanese Monosyllables," J. Radio Res. Labs., Vol. 10, No. 49, 1963. 

169. Suzuki, I., Nakata, K. , and K. Maezono, "Speech Data Analyzed by 

Computer Program, Classification and Recognition of Japanese Mono- 
syllables," IEEE Trans. Military Electronics, MIL-7, No. 2-3, 1963. 

170. Tarneczy, T. H. , "Vowel Formant Bandwidths and Synthetic Vowels," JASA, 

Vol. 34, No. 6, 1962. 

171. Thierney, J. Gold, B. Sferrino, V., Dumanian, I. A., and E. Aho, "Channel 

Vocoder with Digital Pitch Extractor," JASA, Vol.36, No. 10, 1964. 

172. Toffler, I. E. , and F. B. Wade, "Pitch Extractor, Using Clippers," 

JASA, Vol. 36, No. 5, 1964. 

173. Uhr, L. , "Recognition of Letters, Pictures and Speech by a Discovery 

and Learning Program," WESCON Techn. Tapers, Vol. 8, p. 4, August 25- 
28, 1964. 

174. Vilbig, F., "An Analysis of Clipped Speech," JASA, Vol. 27, No. 1, 1955. /399 

175. Vilbig, F., "Speech Compression," JASA 3 Vol. 28, No. 1, 1956. 

176. Vilbig, F. , "Improvement of Simplification of the Scanvocoder and its 

Connection to a Correlation Pulse Code System," J ASA, Vol. 23, No. 4, 
1956. 

177. Vilbig, F. , and K. H. Haase, "Uber einige Systeme zur Sprachbandkompression, 

Nachrichtentechn. Fachber." NTF, Vol. 3, 1956. 

178. Vilbig, F. , and K. H. Haase, "Some Systems for Speech-Band Compression," 

JASA 3 Vol. 28, No. 4, 1956. 

179. Wallace, J. C. , "Comparative Evaluation of Pitch-Signal Indicators," JASA 3 

Vol. 35, No. 5, 1963. 

180. Waller, R. , "Self-programming Pattern Recognizer, Measurements and Control," 

No. 3, 1964. 

181. Waller, R. , "Sceptron," J. Scient. Instruments 3 Vol. 41, No. 5, 1964. 

182. Weiss, M. R. , and C. M. Harris, "Computer Technique for High-Speech 

Extraction of Speech Parameters," JASA, Vol. 35, No. 2, 1963. 

183. Widrow, B. , Groner, G. F., Hu, M. I. C. , Smith, F. W. , Specht, D. F., and 

L. R. Talbert, "Practical Applications for Adaptive Data-Processing 
Systems," WESCON Techn. Papers 3 Vol. 7, No. 7, 1963. 

184. Winckel, F. , "Tonhohenextractor fur Sprache mit Gleichstromanzeige," 

Phonetica, Nr. 3-4, 1964. 

185. Wiren, I., and H. L. Stubbs, "Electronic Binary Selection System for 

Phoneme Classification," JASA 3 Vol. 28, No. 6, 1956. 

186. Wood, D. E. and T. L. Hewitt, "New Instrumentation for Making Spectro- 

graphs Pictures of Speech," JASA, Vol. 35, No. 8, 1963. 

187. Wood, D. E. , "New Display Formant and a Flexible-Time Integration for 

Spectral-Analysis Instrumentation," JASA, Vol. 36, No. 4, 1964. 

188. W. S. , Dr., "Identification des personnes par la spectrographie vocale," 

Automatisme , Vol. 8, No. 7-8, 1963. 

189. Yoshima, T. , "Japan Firm Builds Spoken Voice Digit Recognizer," Electronic 

News, Vol. 8, No. 390, 1963. 

190. Zwicker, E. , "Moglichkeiten zur Spracherkennung uber den Tastsinn mit 

Hilfe eines Funktionsmodells des Gehors," Elektronische Rechenaniagen, 
Vol. 6, H. 6, 1964. 



32